Position-Specific HIV Risk in a Large Network of Homeless Youths

Abstract

Objectives. We examined interconnections among runaway and homeless youths (RHYs) and how aggregated network structure position was associated with HIV risk in this population.

Methods. We collected individual and social network data from 136 RHYs. On the basis of these data, we generated a sociomatrix, accomplished network visualization with a “spring embedder,” and examined k-cores. We used multivariate logistic regression models to assess associations between peripheral and nonperipheral network position and recent unprotected sexual intercourse.

Results. Small numbers of nominations at the individual level aggregated into a large social network with a visible core, periphery, and small clusters. Female youths were more likely to be in the core, as were youths who had been homeless for 2 years or more. Youths at the periphery were less likely to report unprotected intercourse and had been homeless for a shorter duration.

Conclusions. HIV risk was a function of risk-taking youths' connections with one another and was associated with position in the overall network structure. Social network–based prevention programs, young women's housing and health programs, and housing-first programs for peripheral youths could be effective strategies for preventing HIV among this population.

The nearly 2 million runaway and homeless youths (RHYs) in the United States each year¹ are at increased risk for contracting HIV and other sexually transmitted infections.^2–4 Many RHYs are sexual minority youths, and most RHYs come from disorganized family environments filled with conflict, violence, and parental substance abuse.^5–10 When youths run away or are thrown out of the home, they often become embedded in small networks composed of other homeless peers who engage in high-risk behaviors, which normalizes heightened HIV risk taking, such as unprotected sexual intercourse and injection drug use.^5–14 This finding was based on studies of small qualitative samples or egocentric network studies (limited to perceived social ties of the surveyed individual)^5–14 rather than sociometric data that provide linkages among respondents in a sample. School-based sociometric data, however, have shown that a youth's position in broader, larger networks affects health (e.g., suicidal ideation, sexual risk, and smoking).^15–23

Few studies have examined how HIV risk behaviors are tied to sociometric position in larger social network structures, not just among RHYs but also among high-risk youths in general.²⁴ The shift in perspective from specific peer relationships to large social networks requires sociometric data and sociometric theorizing. On the basis of egocentric observations, Whitbeck²⁵ and Rice et al.^26–29 have suggested that there may be 2 distinct types of peer engagement among RHYs. Some youths become embedded in networks of other high-risk homeless youths, whereas other youths, who engage in relatively healthier behaviors, may never become fully embedded in such networks and instead maintain prosocial ties to home, primarily via the Internet and cell phones.^28,29

To move beyond egocentric findings that link risk-taking individuals to risk-taking peers, one must theoretically delineate the ways in which social network position can be related to HIV risk taking. In so doing, one must consider how macro-level social structures (such as housing) may affect the arrangement of social network ties for these youths. Previous work suggests that as youths are homeless for longer periods, they become increasingly likely to form relationships with other homeless youths.^{5–14,25–29} Figure 1 provides a graphical representation of 4 possible arrangements of ties in which risk-taking youths are adjacent to one another and in which the average number of ties per youth is 3; however, the aggregate network structures and locations of risk are quite distinct among the arrangements.

FIGURE 1—

Four equally plausible hypothetical networks—(a) clusters, (b) chain, (c) risk at the core, and (d) risk at the periphery—in which the average degree is 3, one third of youths are risk takers, and risk clusters from an egocentric perspective.

Note. Circles indicate youths; lines indicate social network ties; gray indicates risk taker; white indicates non–risk taker.

The small number of ties reported by RHYs might aggregate into a large number of relatively disconnected networks, or “clusters.” HIV risk occurs in isolated clusters (probably clusters comprising youths who have been homeless longer). Another possibility is that the small number of ties aggregate into long chains of relations that only connect 1 or 2 youths at a time but that span a large number of youths in the aggregate. Again, one might anticipate that risk occurs in particular clusters, primarily among youths who have been homeless longer.

The third and fourth possibilities conceptualize ties as aggregating into large networks that have a core and periphery. In 1 case, youths in the core could be engaged in more risk taking. This scenario presents an interesting tension. The RHYs who have been homeless for longer periods of time (and who are thus relatively more marginal to housed society) become more central to street-based social networks. This marginalization may increase their affiliation with risk-taking peers and may further alienate them from prosocial influences, increasing their chances of engaging in a variety of risk behaviors. Conversely, risk could occur among members of the periphery, and youths who have been homeless for longer may be less well connected to one another (although this possibility seems less likely).

To date, the data needed to assess network structures such as these have not been collected because RHYs are an unbounded population, and social network researchers have not reached consensus on how to collect sociometric data from unbounded populations. The event-based approach (EBA) proposed by Freeman and Webster³⁰ that samples the regulars of a sociophysical space (e.g., a beach) seems promising for RHYs. Freeman and Webster were interested in sampling from natural settings; that is, places where regular patterns of ongoing relationships occur and where entry and exit are common. In such settings, clusters of individuals interact over time, but strictly defined, mutually exclusive categories of individuals cannot be observed in any direct way.³⁰ Freeman and Webster tested this method with regular members of a beach community. They included in their sample those persons who attended the beach 3 times or more in the previous month, thus eliminating transient attendees who were unlikely to be a meaningful part of the social structure at that time. The beach was thus defined as a sociophysical space where interactions occurred among regular attendees over a specified time.

Regular attendance at homeless youth drop-in centers fulfills the requirements of the EBA of Freeman and Webster.³⁰ Drop-in centers, which are common in urban settings, are physical locations where youths congregate to access services such as food, clothing, and case management, but they are also safe havens where youths can freely interact under the loose supervision of adult staff who ensure the safety and peace of the drop-in centers. Drop-in centers exhibit the basic criteria of the natural settings for which the technique of Freeman and Webster technique is appropriate: (1) drop-in centers are a natural social setting where patterns of ongoing social networks may be observed (i.e., a specific sociophysical space); (2) clusters of youths interact in these centers over time, but strictly defined, mutually exclusive categories of individuals are rare; and (3) entry into and exit from these spaces is routine and unregulated.

We used the EBA strategy to record the social ties among a sample of 136 RHYs who received services multiple times from a single drop-in center. We used network visualization techniques³⁰ to examine the resulting network structure, and we performed a k-core analysis.^31,32 We then assessed the associations among relevant network positions and recent condom use.

METHODS

In 2008, we recruited an EBA³⁰ sample of 136 adolescents in Los Angeles, California, at a drop-in agency serving homeless adolescents. All clients aged 13 to 24 years receiving services at the agency during the period of study were eligible to participate, as long as they had received agency services on at least 1 previous occasion. When adolescents signed up to receive services at the agency (e.g., a shower, clothing, case management), they were asked if they would like to participate in the survey. Only 14 (9.3%) of the 150 youths who were asked to participate declined. A consistent set of 2 research staff members was responsible for all recruitment, to prevent adolescents from completing the survey multiple times.

The EBA is not the only viable solution for sampling unbounded populations, but it is the most appropriate strategy for collecting data on network influences on health risk behaviors among homeless youths. In a recent review, Marsden outlined 3 generic boundary specification strategies.^33,34 Marsden's positional approach is based on delineating formal membership in some community or organization and sampling from that list. This approach is typically used in sociometric studies of adolescents in school settings, where classroom rosters or other formal membership lists are available. Although this technique could work well for shelter programs, street-living RHYs tend to be too transient and to lack enduring attachments, making this technique problematic for that population. Second, the relational approach is based on defining the key social relation of interest and sampling individuals who share that relation. Heckathorn's much-used respondent-driven sampling is an example of such a technique and is very useful for finding hidden populations.³⁵ Another approach that has received attention is the expanding selection model proposed by Doreian and Woodard,³⁶ which starts with individuals on some fixed list or agency roster and expands selection to include ties nominated by a minimum number of people in the existing sample. Because both of these techniques sample through relationships, social isolates are excluded, and those who are peripheral (i.e., who have small numbers of ties to others) may be underrepresented. In studying RHYs and HIV risk, we are especially interested in the possibility that relatively healthier behaviors may be manifest among isolated and peripheral youths, so the relational approach does not serve us well. Third, the EBA,³⁰ which bounds individuals on the basis of participation in a shared set of activities or events over time, seems the most applicable to RHYs. The EBA creates a boundary within which to sample youths, does not depend on specific membership in a formal group, and allows social isolates and peripheral youths to be equally represented as highly interconnected youths.

Network Data Collection and Construction

We collected behavioral data from self-administered computer-assisted interviews. All demographic and behavioral variables were based on self-reports (Table 1). Condom use was dummy coded based on an item developed for the Centers for Disease Control and Prevention's Youth Risk Behavior Survey, which has been rigorously tested for reliability and validity^37–39: “The last time you had sexual intercourse, did you or your partner use a condom?”

TABLE 1—

Descriptive Characteristics of Homeless Youths (n = 136): Los Angeles, CA, 2008

Characteristic	No. (%)
Gender
Female	53 (39.6)
Male	81 (60.5)
Sexual orientation: gay, lesbian, or bisexual	17 (12.5)
Race/ethnicity
African American	48 (38.1)
Latino	31 (24.6)
White	27 (21.4)
Mixed race/ethnicity and other	20 (15.9)
Homeless ≥ 2 years	90 (66.2)
Current living situation: literal homelessness
Street	42 (31.1)
Shelter	24 (17.8)
Hotel, motel	10 (7.4)
Current living situation: unstable housing or “couch surfing”	60 (43.7)
Health risk behaviors: condom use at most recent sexual intercourse
Not sexually active previous 90 d	28 (20.6)
Used condom	49 (36.0)
Unprotected	59 (43.4)
Network properties: k-core membership
0-core	37 (27.2)
1-core	40 (29.4)
2-core	24 (17.7)
3-core	4 (2.9)
4-core	7 (5.2)
5-core	2 (1.5)
6-core	9 (6.6)
7-core	13 (9.6)

Note. Mean age = 20.8 years (SD = 2.1). Density Δ = L/[g(g-1)] (where L = ties, g = possible ties) = 0.016.

A trained interviewer collected network data in a face-to-face interview. The following text was read aloud:

Think about the last month. Now I am going to draw a map of your network. We are interested in the people you interact with. We're interested in the people you talk to, “hang out”/“kick it”/“chill” with, have sex or “hook up” with, “party,” drink or use drugs with.

The interviewer then read the following set of prompts to elicit network connections (the wording used was based on how adolescents talk to one another for understanding and clarity), allowing time for reflection and recording:

Friends, family, people you “hang out”/“chill”/“kick it”/have conversations with, people you “party” with—use drugs or alcohol, boyfriend/girlfriend, people you are having sex or “hook up” with, “baby mama”/“baby daddy,” case worker or agency staff, people from school, people from work, old friends from home, people you talk to (on the phone, by e-mail), people from where you are staying (“squatting with”), people you see at this agency, other people you know in Hollywood.

After adolescents finished nominating persons, the following attributes of each nomination were collected: first name and last initial, aliases, age, gender, race/ethnicity, and whether the nominee was a client of the agency.

We created a sociomatrix linking participants in the sample. A directed tie from participant i to participant j was recorded if participant i nominated participant j in his or her personal network. Matches were based on name, alias, ethnicity, gender, approximate age, and agency attendance. If 2 distinct youths matched on all information, presence of a third common tie in each personal network was used to assign adjacency. Questionable matches were left uncoded; hence, we constructed a conservative matrix of ties. Two research assistants each created independent adjacency matrices. The matrices were combined, and discrepant ties were dropped. Only 2 ties were discrepant (99.99% agreement across 18 360 possible ties). The resulting matrix had 288 directed ties (Figure 2).

FIGURE 2—

Social network connections among homeless youths (n = 136): Los Angeles, CA, 2008.

Note. RHY = runaway and homeless youths; circles = homeless  ≥ 2 years; diamonds = homeless  < 2 years. Arrows indicate direction of nominations between youths; gray = unprotected sexual activity at most recent sexual intercourse; white = condom users or not sexually active in previous 90 days.

Network Visualization

We entered data into NetDraw Graph Visualization Software version 2.090 (Analytic Technologies, Irvine, CA) and used the spring embedder routine to generate the network visualizations presented in Figure 2. Spring embedding is based on the idea that 2 actors may be thought of as pushing or pulling each other; 2 points located close together represent actors who have a pull on each other, and distant actors push one another apart. The algorithm seeks a global optimum where there is the least stress on the “springs” connecting actors to one another.⁴⁰

Network Analysis

Because the network shows a clear core, with a periphery that included a large number of isolates, we examined k-cores with respect to HIV risk. A k-core is a subgraph in which each node is adjacent to a minimum of k of the other nodes.²⁶ For this network, k-cores 0 through 7 can be assigned. Periphery membership was defined by k-core 0 or 1 indicating that a youth was either an isolate or had only 1 tie to another network member. We ran a multivariate logistic regression model to assess individual-level covariates with core membership (relative to periphery).

Statistical Analysis

Although recent work has begun to successfully integrate social network analytic techniques with statistical models of individual-level behaviors, there is no agreed-upon best method for conducting these analyses. Sociometric data violate the assumptions of independent observations, which underlie linear statistics. For this analysis, we examined k-cores because of the network visualization but also because once k-core membership is derived from the network model, it can be assigned to individuals as an attribute and incorporated into individual-level statistical models. We ran bivariate logistic regression models assessing associations between k-core membership, time spent homeless, and unprotected sexual intercourse. Based on these preliminary associations, we created multivariate logistic regression models to assess associations among individual characteristics, network position, and HIV risk.

RESULTS

As shown in Table 1, there were slightly more males than females included in the sample. Most RHYs were racial/ethnic minorities. Fifty-six percent were literally homeless, and the remainder were unstably housed, many “couch surfing” (i.e., temporarily staying with friends or relatives). Fifty-seven percent had a lifetime history of using cocaine, heroin, or methamphetamines. With respect to sexual risk, 43% of the sample reported recent unprotected sexual intercourse.

Figure 2 depicts the network of 136 RHYs. Small numbers of ties aggregate into larger network structures of RHYs. Visual inspection of this network reveals a core, a periphery, and a large number of isolates. Table 1 presents the distribution of RHYs with respect to cores. k-cores 0 and 1—that is, youths who were either isolates or had only 1 tie to other youths in the network—constituted 56.6% of the network. We labeled these youths the periphery, and we labeled youths in k-cores 2 through 7 nonperipheral.

The bivariate associations presented in Table 2 show that position (as defined by k-core) is associated with time spent homeless and health risk. Nonperipheral youths were 3.6 times more likely to report having been homeless for 2 or more years. These nonperipheral youths were also more likely to report recent unprotected sexual intercourse (odds ratio [OR] = 2.2; 95% confidence interval [CI] = 1.1, 4.4).

TABLE 2—

Bivariate Logistic Regression of Network Properties, Time Spent Homeless, and HIV Risk Behaviors: Homeless Youths, Los Angeles, CA, 2008

	Homeless ≥ 2 Years, OR (95% CI)	Recent Unprotected Sexual Intercourse, OR (95% CI)
Homeless ≥ 2 y		1.72 (0.82, 3.58)
Nonperipheral (k-core 2–7)	3.64** (1.64, 8.05)	2.20* (1.10, 4.39)

Note. CI = confidence interval; OR = odds ratio. Sample size was n = 136.

*P < .05; **P < .01.

The multivariate models presented in Table 3 show that males were 3.0 times more likely to have been in the periphery of this network (k-cores 0 or 1) and that youths being homeless for 2 or more years was associated with a 72% reduction in the odds of being on the periphery. Nonperipheral youths (k-cores 2 through 7) were 2.2 times more likely to report that their most recent sexual intercourse was unprotected after we controlled for individual-level covariates, including time spent homeless.

TABLE 3—

Multivariate Logistic Regression of k-Core Membership and Time Spent Homeless on Unprotected Sexual Intercourse: Homeless Youths, Los Angeles, CA, 2008

	Network Periphery Member,a OR (95% CI)	Recent Unprotected Sexual Intercourse, OR (95% CI)
Male	3.03** (1.40, 6.55)	0.68 (0.31, 1.49)
African American	0.94 (0.43, 2.07)	1.04 (0.48, 2.26)
Age	1.07 (0.89, 1.29)	1.19 (0.99, 1.42)
Gay, lesbian, bisexual	1.72 (0.51, 5.86)	3.06 (0.94, 9.96)
Homeless 2+ years	0.28** (0.12, 0.67)	0.93 (0.40, 2.15)
Non-peripheral (k-core 2–7)		2.21* (1.01, 4.85)
−2 log likelihood	161.64	169.03

Note. CI = confidence interval; OR = odds ratio. Sample size was n = 136.

^aNetwork periphery defined as k-cores 0 and 1.

*P < .05; **P < .01.

DISCUSSION

Although it has been suggested that homeless youths have relatively small social networks,^11–14 these data show how even a small number of ties to other youths aggregates to a large social network. Although peer relationships have held a central role in theoretical and empirical work on homeless youths, sociometric data such as these depicting the larger web of interconnected ties have never before been collected. Collecting sociometric data like these on unbounded populations is both conceptually and logistically difficult. We used the EBA technique of Freeman and Webster³⁰ and successfully captured isolates and youths on the periphery, as well as highly interconnected youths.

This network structure, consisting of 1 large interconnected core, several smaller clusters, and many isolates, may be common among risk-taking youths in other populations. Friedman et al. found the same overall network structure in their Network Norms and HIV Risk Among Youth project.²⁴ Logically, this is a hybrid of 2 of many possible structures (Figure 1), but because 2 different high-risk youth populations can be characterized by this arrangement of ties, this may be a typical, or at least recurring, structure.

These data reveal that position in larger interconnected webs of other RHYs affects the HIV risk-taking behaviors of RHYs. Although these data are preliminary, they suggest that peripheral youths were less engaged in sexual risk behaviors. Research on homeless youths has shown for quite some time that for a particular RHY, HIV risk behaviors are strongly associated with having peers who engage in these risky behaviors. The affiliations among risk-taking youths could be the result of social learning or selection processes.¹³ Regardless, these data suggest that where a particular youth is located in the larger structure of homeless youths is associated with risk, as opposed to just having a risk-taking peer. Peripheral locations are lower risk than core locations. This contradicts studies from high-school settings that suggest the positive aspects of social integration.^15–23 Here, being peripheral to a risk-taking network is associated with lowered risk. If this structure can be generalized to other high-risk youth populations, then perhaps peripheral positions in those networks will also be associated with lowered risk.

These data suggest that social network position and macro-level structures—large-scale social phenomena such as poverty, housing, racism, sexism, and homophobia⁴¹—are linked. Peripheral position (network structure) was associated with male gender and shortened duration of homelessness. Peripheral position, in turn, was associated with reduced HIV risk. The greater number of females at the core may reflect larger societal structures that make females much more vulnerable to physical coercion and sexual exploitation, which are common in street life,^9,10 increasing their dependency on other homeless males. Position and duration homeless hint at a process of social network formation whereby newly homeless youths start out at the periphery of the network and become embedded over time.^25–29 Youths at the periphery, who have been homeless for less time and who take fewer risks, may still have attachments to macro-social structures such as home and school.^25–29

This study has a few limitations that must be acknowledged. First, these data are cross-sectional, so causality cannot be determined; nor can we assert the existence of a process taking place over time. Second, this preliminary attempt to use the EBA approach is imperfect. Although we enrolled as many repeat visitors as possible and had a relatively low refusal rate, we did not capture the entire population of repeat visitors during the period under observation. Moreover, these data come from only 1 agency, and despite the diversity of homeless youths in our sample (street, shelter, and “couch surfing”), we cannot say that these data are generalizable to homeless youths in general.

Despite the preliminary nature of these data, our findings have 3 important implications for future research. First, Freeman and Webster's EBA seems a very promising technique for sampling an unbounded population who regularly attend some sociophysical space such as a drop-in center for homeless youths. This approach captures youths who are isolates, peripheral, and highly connected alike. Developing increasingly rigorous procedures for collecting such samples and collecting these data would do much to reveal the causal processes at work in risk-taking behaviors among homeless youths.

Second, these data begin to give us some direction for what kinds of peer-led programs would be most effective for RHYs. Arnold and Rotheram-Borus⁴² recently reviewed the body of effective HIV prevention interventions for RHYs and concluded that the “next generation” of programs ought to focus on network-based models, because network-based programs cost less and are more easily disseminated in community settings. These data provide some direction regarding how to proceed. Although many of these youths are connected, more than half of the network can be considered isolates or peripheral. Moreover, the highest-risk youths are also the most frequently nominated (i.e., most popular). As a result, the Popular Opinion Leader model,⁴³ which relies on recruiting the core, seems less viable than a model like SHIELD,⁴⁴ which looks to make change across a network via localized action around a dispersed group of peer leaders located throughout a risk-taking network.

Third, these data suggest 2 connections between network-level phenomena and macro-level factors that have implications for interventions. Female homeless youths were found disproportionately at the core of this street-based network where HIV risk was elevated, suggesting that female-specific programs for young women's health and housing may be needed to enable these young women to extricate themselves from these networks. Although housing is a desired outcome for all homeless youths, many homeless youths are not socially or psychologically able to participate in housing programs.¹⁰ Housing interventions, especially housing-first interventions, may have added traction for youths at the periphery of street-based networks, because they have been homeless for a shorter time, and street life may exert a weaker pull on them. These youths, who are socially peripheral to street life, may be in an advantaged position with respect to their capacity to engage with the structures and opportunities provided by housing programs.